JP7315339B2 - Time-series data division method, division device, and program - Google Patents

Description

The present invention relates to a time-series data division method, a division device, and a program.

Conventionally, in order to extract a pattern included in time-series data, work is performed to divide the time-series data so that each portion includes the pattern. Each part of the divided time-series data is used as teacher data or the like for learning patterns by machine learning.

JP 2018-147406 A

In general, division of time-series data is performed visually by an operator. Therefore, when the data length of the time-series data is long, or when there are a large number of files of the time-series data, there is a problem that the dividing work takes a lot of time and effort.

In order to automate the division of time-series data, a method of automatically dividing time-series data according to division rules has also been proposed. However, in this method, it is necessary to set division rules in advance according to the characteristics of time-series data, and it takes time to set division rules.

The present invention has been made in view of the above problems, and an object of the present invention is to enable easy division of time-series data.

A division method according to one embodiment is a method for dividing time-series data executed by an information processing apparatus, and includes an extraction process for extracting a feature amount from the time-series data, a division process for dividing the time-series data by a plurality of division methods based on the feature amount, an evaluation process for calculating an evaluation value for each of the division methods based on the divided time-series data, and a selection process for selecting at least one division method from the plurality of division methods based on the evaluation value.

According to each embodiment of the present invention, time-series data can be easily divided.

The figure which shows an example of the hardware constitutions of a division|segmentation apparatus. The figure which shows an example of the functional structure of a division|segmentation apparatus. 4 is a flowchart showing an example of a method for dividing time-series data; The figure which shows an example of time-series data. FIG. 5 is a diagram showing an example of trend data generated from the time-series data of FIG. 4; FIG. 6 is a diagram showing an example of feature amounts extracted from the trend data of FIG. 5; The figure which shows an example of the time series data of FIG. 4 which were divided|segmented. A table showing an example of evaluation values of division methods. The figure which shows an example of the screen which displays a division|segmentation position.

Hereinafter, each embodiment of the present invention will be described with reference to the accompanying drawings. In addition, regarding the descriptions of the specifications and drawings according to each embodiment, components having substantially the same functional configuration are denoted by the same reference numerals, thereby omitting overlapping descriptions.

A time-series data dividing device 1 according to an embodiment will be described with reference to FIGS. 1 to 9. FIG. A splitting device 1 according to the present embodiment is an information processing device (computer) that automatically performs splitting processing of time-series data. In this embodiment, it is assumed that time-series data including a periodic pattern is divided so that each portion includes the pattern. The time-series data is, for example, sensor data, but is not limited to this.

First, the hardware configuration of the dividing device 1 will be described. FIG. 1 is a diagram showing an example of the hardware configuration of the dividing device 1. As shown in FIG. 1 includes a CPU (Central Processing Unit) 101, a ROM (Read Only Memory) 102, a RAM (Random Access Memory) 103, a HDD (Hard Disk Drive) 104, an input device 105, a display device 106, a communication interface 107, and a bus 108.

The CPU 101 executes programs to control each component of the dividing device 1 and implement the functions of the dividing device 1 . A program executed by the CPU 101 can be recorded on any computer-readable recording medium such as a CD (Compact Disk), DVD, flash memory, or the like.

The ROM 102 stores programs executed by the CPU 101 and various data.

A RAM 103 provides a work area for the CPU 101 .

The HDD 104 stores programs executed by the CPU 101 and various data. The partitioning device 1 may include an SSD (Solid State Drive) together with the HDD 104 or instead of the HDD 104 .

The input device 105 inputs information according to the user's operation to the dividing device 1 . The input device 105 is, for example, a touch pad, keyboard, mouse, or hardware buttons, but is not limited thereto.

The display device 106 displays a screen according to the user's operation. The display device 106 is a liquid crystal display or an organic EL (Electro Luminescence) display, but is not limited to this. The splitting device 1 may include a touch panel in which the touch pad and the display device 106 are integrated.

The communication interface 107 connects the dividing device 1 to a network such as the Internet or a LAN (Local Area Network). The splitting device 1 communicates with external devices on the network via the communication interface 107 .

A bus 108 connects the CPU 101, ROM 102, RAM 103, HDD 104, input device 105, display device 106, and communication interface 107 to each other.

Note that the hardware configuration of the dividing device 1 is not limited to the example in FIG. For example, the dividing device 1 may be realized by one information processing device as in the example of FIG. 1, or may be realized by a plurality of information processing devices connected via a network such as the Internet or LAN.

Next, the functional configuration of the dividing device 1 will be described. FIG. 2 is a diagram showing an example of the functional configuration of the dividing device 1. As shown in FIG. The dividing device 1 of FIG. 2 includes an extraction unit 11 , a dividing unit 12 , an evaluating unit 13 , a selecting unit 14 and a display unit 15 . These functional configurations are realized by the CPU 101 executing programs and cooperating with other hardware configurations.

The extraction unit 11 extracts one or more feature quantities from the time-series data D1. The feature amount includes, but is not limited to, each time-series value calculated by the autocorrelation function (hereinafter referred to as "autocorrelation value"), each time-series value calculated by the cross-correlation function (hereinafter referred to as "cross-correlation value"), each time-series value calculated by change amount calculation (hereinafter referred to as "change amount"), and each time-series value calculated by dynamic time warping (DTW) (hereinafter referred to as "DTW value"). The extraction unit 11 may directly extract the feature amount from the time-series data D1, or may generate trend data D2 from the time-series data D1 and extract the feature amount from the generated trend data D2. The trend data D2 is data obtained by preprocessing the time-series data D1 so as to facilitate the extraction of feature amounts. Preprocessing includes, but is not limited to, filter processing, moving average processing, and convolution processing, for example. Filtering includes low-pass filtering, high-pass filtering, and band-pass filtering.

The division unit 12 divides the time-series data D1 by a plurality of division methods based on the feature amount of the time-series data D1 extracted by the extraction unit 11 . In other words, the division unit 12 determines division positions p of the time series data D1 by a plurality of division methods based on the feature amount of the time series data D1 extracted by the extraction unit 11 .

The dividing unit 12 may divide the time-series data D1 by one division method based on one feature amount, or may divide the time-series data D1 by a plurality of division methods based on one feature amount. The dividing unit 12 may divide the time-series data D1 at the peak P of the feature amount, for example. In this case, the position of peak P corresponds to division position p. Further, the dividing unit 12 may divide the time-series data D1 at peaks P equal to or greater than a preset threshold. Note that the method of dividing the time-series data D1 is not limited to the above example.

Based on the divided time-series data D1, the evaluation unit 13 evaluates the dividing method for dividing the time-series data D1. Specifically, the evaluation unit 13 calculates an evaluation value s for each division method. The evaluation value s is an index indicating the appropriateness of the dividing method. In the following, an example will be described in which the larger the evaluation value s, the more appropriate the division method. However, the magnitude relationship between the evaluation value s and appropriateness may be reversed.

The evaluation value s is calculated based on the degree of similarity between the divided parts of the time-series data D1. Examples of similarity measures include, but are not limited to, spectral distances, cross-correlation values, and DTW values. For example, when the evaluation value s is a cross-correlation value and the time-series data D1 is divided by the division method X, the evaluation unit 13 calculates the cross-correlation value between the parts of the time-series data D1 divided by the division method X as the evaluation value s of the division method X.

Also, the evaluation value s may be calculated based on a plurality of degrees of similarity. For example, when the evaluation value s is the sum of the spectral distance and the cross-correlation value, and the time-series data D1 is divided by the division method X, the evaluation unit 13 calculates the spectral distance and the cross-correlation value between the parts of the time-series data D1 divided by the division method X, normalizes the calculated spectral distance and the cross-correlation value, and calculates the sum of the normalized spectral distances and the cross-correlation values as the evaluation value s of the division method X.

Note that the method of calculating the evaluation value s is not limited to the above example. The evaluation value s may be the sum of three or more similarities, the average of multiple similarities, or the weighted sum of multiple similarities.

Based on the evaluation value s, the selection unit 14 selects the division method with the largest evaluation value s from among the plurality of division methods. The selection unit 14 may select a plurality of division methods in descending order of the evaluation value s.

The display unit 15 displays the division position p of the time-series data D1 by the division method selected by the selection unit 14 on the display device 106 together with the time-series data D1. The display unit 15 preferably displays the division position p so that the user can adjust the displayed division position p by operating the input device 105 . Thereby, the user can finely adjust the division position p of the time-series data D1.

Next, a method of dividing the time-series data D1 executed by the dividing device 1 will be described. FIG. 3 is a flow chart showing an example of a method of dividing the time-series data D1.

First, time-series data D1 is input to the dividing device 1 (step S101). FIG. 4 is a diagram showing an example of the time-series data D1. In the example of FIG. 4, the time-series data D1 includes four predetermined patterns.

Next, the extraction unit 11 generates trend data D2 from the input time-series data D1 (step S102). FIG. 5 is a diagram showing an example of trend data D2 generated from the time-series data D1 of FIG. The trend data D2 in FIG. 5 is generated by applying a low-pass filter to the time-series data D1 in FIG. Note that, as described above, the filtering process is not limited to a low-pass filter.

Subsequently, the extraction unit 11 extracts feature amounts from the generated trend data D2 (step S103). FIG. 6 is a diagram showing an example of feature amounts extracted from the trend data D2 of FIG. The feature quantity in FIG. 6 is the cross-correlation value extracted from the trend data D2 in FIG.

When the feature amount of the time series data D1 is extracted, the division unit 12 divides the time series data D1 based on the feature amount (step S104). In the example of FIG. 6, it is assumed that the dividing unit 12 determines, among the peaks P of the feature amount, the peaks P that are equal to or greater than a preset threshold value as the dividing positions p. When the preset threshold is the threshold Th1, the peak P1 is determined as the division position p, and the time series data D1 is divided at the peak P1. When the preset threshold is the threshold Th2, the peaks P1 and P2 are determined as the division position p, and the time series data D1 is divided at the peaks P1 and P2. When the preset threshold is the threshold Th3, the peaks P1 to P3 are determined as the division position p, and the time series data D1 is divided at the peaks P1 to P3.

FIG. 7 is a diagram showing an example of the time-series data D1 of FIG. 4 divided by the dividing unit 12. As shown in FIG. In the example of FIG. 7, the time-series data D1 of FIG. 4 is divided by peaks P1 to P3 of FIG. That is, the time-series data D1 in FIG. 4 is divided into a portion d1 before the peak P1, a portion d2 from the peak P1 to the peak P2, a portion d3 from the peak P2 to the peak P3, and a portion d4 after the peak P3.

In the above example, it is assumed that a plurality of division positions p are determined simultaneously, but division positions p may be determined one by one. In this case, after determining one division position p (e.g., peak P1), the feature quantity is again extracted for the time-series data D1 after the division position p, and one division position p is determined again based on the obtained feature quantity. By repeating extraction of the feature amount and determination of the division position p until the division position p cannot be determined, the entire time-series data D1 can be divided.

After dividing the time-series data D1 by the dividing unit 12, the evaluating unit 13 calculates an evaluation value s of the dividing method executed by the dividing unit 12 (step S105). For example, when the time-series data D1 is divided into portions d1 to d4 as shown in FIG. 7, the evaluation unit 13 calculates the evaluation value s based on the degree of similarity of the portions d1 to d4.

If division of the time-series data D1 by a plurality of division methods has not been completed (step S106: NO), the process returns to step S102, and the division unit 12 divides the time-series data D1 by a division method different from the previous time. At this time, the extraction unit 11 may generate the trend data D2 by a generation method different from the previous time, or may extract the feature amount by an extraction method different from the previous time. Correspondence among the method of dividing the time-series data D1, the method of generating the trend data D2, and the method of extracting the feature amount is set in advance.

On the other hand, when the division of the time-series data D1 by a plurality of division methods is completed (step S106: YES), the selection unit 14 selects the division method with the largest evaluation value s from among the plurality of division methods based on the evaluation value s of each division method (step S107).

FIG. 8 is a table showing an example of evaluation values s of division methods. In the example of FIG. 8, the dividing unit 12 divides the time-series data D1 by three dividing methods A to C, and the evaluating unit 13 calculates six similarities 1 to 6 for each dividing method, and the sum of the similarities 1 to 6 is calculated as the evaluation value s. In the example of FIG. 8, the division method A is selected as the division method with the largest evaluation value s.

When the dividing method is selected by the selecting unit 14, the display unit 15 displays the dividing position p determined by the selected dividing method together with the time-series data D1 on the display device 106 (step S108).

FIG. 9 is a diagram showing an example of the screen W displaying the division position p. Screen W in FIG. 9 corresponds to screen W when division method A is selected based on evaluation value s in FIG. The screen W of FIG. 9 displays time-series data D, division positions p1 to p3 determined by the division method A selected by the selection unit 14, similarities a to f and the evaluation value s of the division method A calculated by the evaluation unit 13, switching buttons B1 to B3 for switching the division method displayed on the screen W to the division methods B to D, and an enter button B4 for determining the division position p.

By displaying the screen W as shown in FIG. 9, the user can easily grasp the division positions p1 to p3 of the time-series data D by the division method A and the appropriateness of the division positions p1 to p3 (the evaluation value s of the division method A). Further, the user can arbitrarily move and adjust the division positions p1 to p3 by operating the input device 105. FIG. Further, by operating the input device 105 and selecting the switch button B1, the user can switch the division position p displayed on the screen W to the division position p by the division method B, and switch the similarity and evaluation value s displayed on the screen W to the similarity and evaluation value s of the division method B. This also applies to the switching buttons B2 and B3.

Further, the user can set the dividing position p of the time-series data D to the dividing position p displayed on the screen W by operating the input device 105 to select the enter button B4. In the example of FIG. 9, when the decision button B4 is selected, the division position p of the time-series data D is determined as division positions p1 to p3. When the division position p of the time-series data D is determined by the user, the determined division position p and each portion d obtained by dividing the time-series data D at the division position p are stored in the HDD 104 . Each portion d thus saved can be used as training data for pattern learning by machine learning.

As described above, the dividing device 1 according to this embodiment can automatically divide arbitrary time-series data D. FIG. As a result, work such as setting division rules according to the time-series data D becomes unnecessary, so that the time-series data D can be easily divided.

Further, the dividing device 1 divides the time-series data D by a plurality of division methods, evaluates each division method, displays the division position p of the time-series data D by the most appropriate division method on the screen W, and can recommend the optimum division position p of the time-series data D to the user.

Note that the dividing device 1 may determine the dividing position p of the time-series data D as the optimum dividing position p without recommending the optimum dividing position p to the user, and automatically divide the time-series data D at the optimum dividing position p.

Further, the splitting device 1 may select a plurality of splitting methods based on the evaluation value s, and display split positions p according to the selected splitting methods on the screen W at the same time.

It should be noted that the present invention is not limited to the configurations shown here, such as combinations with other elements, to the configurations listed in the above embodiments. These points can be changed without departing from the gist of the present invention, and can be determined appropriately according to the application form.

1: Division device 11: Extraction unit 12: Division unit 13: Evaluation unit 14: Selection unit 15: Display unit

Claims (11)

A time-series data division method executed by an information processing device,
an extraction process for extracting a feature amount from the time-series data;
A division process for dividing the time-series data by a plurality of division methods based on the feature amount;
an evaluation process of calculating an evaluation value for each of the division methods based on the divided time -series data, and calculating the evaluation value based on a degree of similarity between portions of the divided time-series data;
a selection process of selecting at least one division method from among the plurality of division methods based on the evaluation value;
A method for segmenting time series data with 2. The method of dividing time-series data according to claim 1 , wherein the similarity includes at least one of spectral distance, cross-correlation value, and DTW value. 2. The time-series data division method according to claim 1 , wherein the extraction process generates trend data from the time-series data, and extracts the feature amount from the generated trend data. The extraction process generates trend data from the time-series data, extracts the feature amount from the generated trend data, and generates the trend data by at least one of filtering, moving average, and convolution.
The method for dividing time-series data according to claim 1. The dividing process divides time-series data including a periodic pattern so that each portion includes the periodic pattern.
The method for dividing time-series data according to claim 1. 2. The method of dividing time-series data according to claim 1, wherein the feature amount includes at least one of an autocorrelation value, a cross-correlation value, an amount of change, and a DTW value. 7. The time-series data division method according to claim 1, wherein the selection process selects the division method with the largest evaluation value. 8. The time-series data division method according to claim 1, further comprising display processing for displaying division positions of the time-series data according to the selected division method. 9. The time-series data division method according to claim 8, wherein the display processing displays the division position of the time-series data so that a user can adjust the division position of the time-series data. an extraction unit that extracts feature quantities from time-series data;
a division unit that divides the time-series data by a plurality of division methods based on the feature amount;
an evaluation unit that calculates an evaluation value for each of the division methods based on the divided time-series data , and the evaluation value is calculated based on a degree of similarity between portions of the divided time-series data;
a selection unit that selects at least one division method from among the plurality of division methods based on the evaluation value;
A device for dividing time series data comprising: information processing equipment,
an extraction process for extracting feature values from time-series data;
A division process for dividing the time-series data by a plurality of division methods based on the feature amount;
an evaluation process of calculating an evaluation value for each of the division methods based on the divided time -series data, and calculating the evaluation value based on a degree of similarity between portions of the divided time-series data;
a selection process of selecting at least one division method from among the plurality of division methods based on the evaluation value;
A program for executing a method for dividing time series data.

Images